Multicentre validation of a microRNA-based assay for diagnosing indeterminate thyroid nodules utilising fine needle aspirate smears

Abstract

Aims: The distinction between benign and malignant thyroid nodules has important therapeutic implications. Our objective was to develop an assay that could classify indeterminate thyroid nodules as benign or suspicious, using routinely prepared fine needle aspirate (FNA) cytology smears.

Methods: A training set of 375 FNA smears was used to develop the microRNA-based assay, which was validated using a blinded, multicentre, retrospective cohort of 201 smears. Final diagnosis of the validation samples was determined based on corresponding surgical specimens, reviewed by the contributing institute pathologist and two independent pathologists. Validation samples were from adult patients (≥18 years) with nodule size >0.5 cm, and a final diagnosis confirmed by at least one of the two blinded, independent pathologists. The developed assay, RosettaGX Reveal, differentiates benign from malignant thyroid nodules, using quantitative RT-PCR.

Results: Test performance on the 189 samples that passed quality control: negative predictive value: 91% (95% CI 84% to 96%); sensitivity: 85% (CI 74% to 93%); specificity: 72% (CI 63% to 79%). Performance for cases in which all three reviewing pathologists were in agreement regarding the final diagnosis (n=150): negative predictive value: 99% (CI 94% to 100%); sensitivity: 98% (CI 87% to 100%); specificity: 78% (CI 69% to 85%).

Conclusions: A novel assay utilising microRNA expression in cytology smears was developed. The assay distinguishes benign from malignant thyroid nodules using a single FNA stained smear, and does not require fresh tissue or special collection and shipment conditions. This assay offers a valuable tool for the preoperative classification of thyroid samples with indeterminate cytology.

Keywords: DIAGNOSTICS; LABORATORY TESTS; MOLECULAR ONCOLOGY; THYROID; THYROID CANCER.

Conflict of interest statement

Competing interests: Authors affiliated with Rosetta Genomics are full-time employees of the company and/or hold equity in the company, which stands to gain from the publication of this manuscript. One of the authors (A. Shtabsky) is a payed consultant for Rosetta Genomics. The authors from medical/clinical centres have received research funding from the company as part of this and/or other collaborative projects.

Published by the BMJ Publishing Group Limited. For permission to use (where not already granted under a licence) please go to http://www.bmj.com/company/products-services/rights-and-licensing/.

Figures

Figure 1

Assay development. The study was composed of three stages: (1) a discovery phase, (2) training, (3) validation. Discovery studies: (I) An initial set of 96 miRNAs was selected based on their differential expression in benign and malignant samples (53 formalin-fixed paraffin-embedded (FFPE), 73 cell block and 84 fine needle aspirate (FNA) samples) as seen on custom microarray and next generation sequencing (NGS) experiments. (II) This set of miRNAs was then evaluated on FNA smears (n=82) using qRT-PCR and 24 miRNAs were selected for the assay training and validation stages. Training: the final assay classifier was developed and cross validated on an FNA training set (n=375). Validation: the test was validated on a blinded set of 189 indeterminate samples from Bethesda classes III, IV and V for which at least one out of two independent pathologists agreed with the original pathologist regarding the final diagnosis (benign or malignant) of the sample. The results of the test on a subset of validation set samples for which all three pathologists agreed (n=150) were also assessed.

Figure 2

Negative predictive value (NPV) and positive predictive value (PPV) for varying prevalence values. NPV and PPV were calculated, based on the observed sensitivity and specificity in the blinded validation set, for varying prevalence values. Dashed lines: the entire validation set (sensitivity: 85.2%, specificity: 71.9%), solid lines: the agreement subset (sensitivity: 97.5%, specificity: 78.2%). Red line: calculated NPV. Blue line: calculated PPV.

Figure 3

Medullary carcinoma linear discriminant analysis (LDA) step. An LDA classifier based on the expression of hsa-miR-375 is used to differentiate medullary carcinoma samples. All the training medullary carcinoma stained smears and two of the three medullary smears in the test set demonstrate overexpression of hsa-miR-375. Yellow diamonds: malignant non-medullary training samples. Blue squares: benign training samples. Green circles: medullary carcinoma training samples. Red stars: medullary carcinoma validation samples.